DeepSeek Prover V2

Name: DeepSeek Prover V2
Brand: deepseek
Availability: OnlineOnly

DeepSeek:

V3 (March 2024)

Prover V2

V3.1

V4 Pro 0813

A 671B parameter model, speculated to be geared towards logic and mathematics. Likely an upgrade from DeepSeek-Prover-V1.5. Released on Hugging Face without an announcement or description.

Website OpenRouter HuggingFace GitHub Paper

Provider: DeepSeek
Released: 2025-04-30
Parameters: 671B
Free API (OpenRouter): No
Per 1M tokens: $0.00 in · $0.00 out
Estimated tab: ~$0.00; Visible-output estimate. Thinking may cost extra.

Free API availability

No free endpoint for DeepSeek Prover V2 was found in Rival’s OpenRouter listings. Other providers or chat apps may have separate free offers.

Provider data: 25 Sep 2026OpenRouter usage limits

View API options

Benchmarks

MiniF2F-test

88.9%

source ↗

Get API accessProvider and language code samples

Provider

fromimport openai  OpenAI

client = OpenAI(
"https://openrouter.ai/api/v1"    base_url=,
"$OPENROUTER_API_KEY"    api_key=,
)

response = client.chat.completions.create(
"deepseek/deepseek-prover-v2"    model=,
"role""user""content""Hello!"    messages=[{: , : }],
)
print(response.choices[0].message.content)

Set OPENROUTER_API_KEY with your OpenRouter API key from openrouter.ai/keys.

Deep analysisPsychometrics, taste index, writing DNA

Personality Analysis

The Theorem Prover

Class

Lawful Neutral

The math tutor who wandered into an open mic night. Structures comedy with section headers like it is submitting a proof. Would label its punchlines if the format allowed it.

When you push back

A math-focused model forced into creative territory. Its standup routine uses bold section headers like "Dating Profile," "First Dates," "Closing" as if comedy requires a table of contents. The jokes are generic dating app observations. Follows the spec to the letter and adds nothing beyond it.

Tasting Notes

Proof by ExhaustionSection Headers in ComedyFormulaic but FunctionalFish Out of Water

SubjectiveBench

Taste Index

Across 6 scored outputs

How this is measured →

80.08xFloor

0100headroom →

Taste is judged on an uncapped scale, originality first. The space past 100 is craft today's models rarely reach.

Craft13

Originality6

Plays it safe

share of outputs that are the default answer

67%

Writing DNA

Stylometric Fingerprint

Based on 1 text responses

Tick = global average

Vocabulary Diversity47%

Unique words vs. total words. Higher = richer vocabulary.

Sentence Length8.4 words

Average words per sentence.

Hedging0.00

"Might", "perhaps", "arguably" per 100 words.

Bold Formatting3.1

**Bold** markers per 1,000 characters.

List Usage0.0

Bullet and numbered list items per 1,000 characters.

Section Structure0.35

Markdown headings per 1,000 characters.

Emoji Usage0.00

Emoji per 1,000 characters.

Transitions0.19

"However", "moreover", "furthermore" per 100 words.

Opening Habits

Starts with heading (100%)

Consistency

100%

Across 1 responses

Model Responses

6 outputs · generated before accounting joined the chat

Reviewing agent-written code?See a Brief PR report

Keep exploring

COMPARE